Response Theory for Equilibrium and Non-Equilibrium ... · Response Theory for Equilibrium and...

22
arXiv:0710.0958v1 [cond-mat.stat-mech] 4 Oct 2007 Response Theory for Equilibrium and Non-Equilibrium Statistical Mechanics: Causality and Generalized Kramers-Kronig relations Valerio Lucarini Department of Physics, University of Bologna, Bologna, Italy (Dated: February 13, 2013) Abstract We consider the general response theory recently proposed by Ruelle for describing the impact of small perturbations to the non-equilibrium steady states resulting from Axiom A dynamical systems. We show that the causality of the response functions entails the possibility of writing a set of Kramers-Kronig relations for the corresponding susceptibilities at all orders of nonlinearity. Nonetheless, only a special class of directly observable susceptibilities obey Kramers-Kronig rela- tions. The apparent contradiction with the principle of causality is also clarified. Specific results are provided for the case of arbitrary order harmonic response, which allows for a very compre- hensive Kramers-Kronig analysis and the establishment of sum rules connecting the asymptotic behavior of the harmonic generation susceptibility to the short-time response of the perturbed system. These results set in a more general theoretical framework previous findings obtained for optical Hamiltonian systems and simple mechanical models, and shed light on the very general im- pact of considering the principle of causality for testing self-consistency: the described dispersion relations constitute unavoidable benchmarks that any experimental and model generated dataset must obey. In order to gain a more complete picture, connecting the response theory for equi- librium and non equilibrium systems, we show how to rewrite the classical response theory by Kubo for systems close to equilibrium so that response functions formally identical to those pro- posed by Ruelle, apart from the measure involved in the phase space integration, are obtained. Finally, we briefly discuss how the presented results, taking into account the chaotic hypothesis by Gallavotti and Cohen, might have relevant implications for climate research. In particular, whereas the fluctuation-dissipation theorem does not work for non-equilibrium systems, because of the non-equivalence between internal and external fluctuations, Kramers-Kronig relations might be more robust tools for the definition of a self-consistent theory of climate change. PACS numbers: 05.20.-y, 05.30.-d, 05.45.-a, 05.70.Ln * Electronic address: [email protected]

Transcript of Response Theory for Equilibrium and Non-Equilibrium ... · Response Theory for Equilibrium and...

arX

iv:0

710.

0958

v1 [

cond

-mat

.sta

t-m

ech]

4 O

ct 2

007

Response Theory for Equilibrium and Non-Equilibrium Statistical

Mechanics: Causality and Generalized Kramers-Kronig relations

Valerio Lucarini∗

Department of Physics, University of Bologna, Bologna, Italy

(Dated: February 13, 2013)

Abstract

We consider the general response theory recently proposed by Ruelle for describing the impact

of small perturbations to the non-equilibrium steady states resulting from Axiom A dynamical

systems. We show that the causality of the response functions entails the possibility of writing a

set of Kramers-Kronig relations for the corresponding susceptibilities at all orders of nonlinearity.

Nonetheless, only a special class of directly observable susceptibilities obey Kramers-Kronig rela-

tions. The apparent contradiction with the principle of causality is also clarified. Specific results

are provided for the case of arbitrary order harmonic response, which allows for a very compre-

hensive Kramers-Kronig analysis and the establishment of sum rules connecting the asymptotic

behavior of the harmonic generation susceptibility to the short-time response of the perturbed

system. These results set in a more general theoretical framework previous findings obtained for

optical Hamiltonian systems and simple mechanical models, and shed light on the very general im-

pact of considering the principle of causality for testing self-consistency: the described dispersion

relations constitute unavoidable benchmarks that any experimental and model generated dataset

must obey. In order to gain a more complete picture, connecting the response theory for equi-

librium and non equilibrium systems, we show how to rewrite the classical response theory by

Kubo for systems close to equilibrium so that response functions formally identical to those pro-

posed by Ruelle, apart from the measure involved in the phase space integration, are obtained.

Finally, we briefly discuss how the presented results, taking into account the chaotic hypothesis

by Gallavotti and Cohen, might have relevant implications for climate research. In particular,

whereas the fluctuation-dissipation theorem does not work for non-equilibrium systems, because

of the non-equivalence between internal and external fluctuations, Kramers-Kronig relations might

be more robust tools for the definition of a self-consistent theory of climate change.

PACS numbers: 05.20.-y, 05.30.-d, 05.45.-a, 05.70.Ln

∗Electronic address: [email protected]

Contents

I. Introduction 3

II. Linear and nonlinear response of perturbed non-equilibrium steady states 5

A. Response of the system in the frequency domain 6

III. Generalized Kramers-Kronig relations 8

A. Basic Results 8

B. A New Definition of Dispersion Relations 9

C. Harmonic Generation 10

IV. Summary and Conclusions 13

Acknowledgments 16

A. Reconciling Kubo’s and Ruelle’s general perturbative response functions 17

B. Response theory for non-equilibrium steady states and climate research 19

References 20

I. INTRODUCTION

The analysis of how systems respond to external perturbations to their steady state

constitutes one of the crucial subjects of investigation in the physical and mathematical

sciences. In the case of physical systems near equilibrium, the powerful approach introduced

by Kubo [1], based on the generalization up to any order of nonlinearity of the formalism of

the Green function, allows for expressing the change in the statistical properties of a general

observable due to the introduction of a perturbation in terms of expectation values of suitably

defined quantities evaluated at the unperturbed state [2]. These results have had huge impact

on statistical mechanics and have allowed detailed treatment of several and rather diverse

processes, including, e.g. the interaction of radiation with condensed matter. Recently,

Ruelle [3, 4] has extended some of the investigations by Kubo to a wide class of systems

far from equilibrium, and introduced a perturbative approach for computing the response

of systems driven away by a small external forcing from their non-equilibrium steady states.

More precisely, the results by Ruelle consider perturbations to autonomous Axiom-A flows

(and maps) defined in a compact manifold, which possess a chaotic, mixing dynamics, and

are associated to an invariant ergodic Sinai-Ruelle-Bown (SRB) measure [5, 6]. We remind

that, the - mathematically speaking, special - case of Axiom A systems amounts to being

of general physical interest, if one accepts the chaotic hypothesis by Gallavotti and Cohen

[7, 8] which states that, for the purpose of computing macroscopic quantities, many-particle

systems behave as though they were dynamical systems with transitive Axiom-A global

attractors. Ruelle shows that, in analogy to what found by Kubo, at all orders of perturbative

expansion, the effect of the forcing on the expectation value of a general observable can be

expressed in terms of averages of quantities performed at the non-equilibrium steady state,

i.e. obtained by integrating over the unperturbed SRB measure. Moreover, in the case of

linear response, it is shown that it is possible to define formally a susceptibility function,

obtained as the Fourier Transform of the linear Green function of the system, and to prove

that such susceptibility, basically as a result of the causality principle, obeys Kramers-

Kronig (K-K) relations [9, 10], just as in Kubo framework. The K-K relations say that

the the real and imaginary part of the linear susceptibility are fundamentally connected,

each one being the Hilbert transform of the other one. Hence, these integral properties

provide unavoidable constraints for checking the self-consistency of experimental or model-

generated data. Furthermore, by applying the K-K relations, it is possible to perform the

so-called inversion of data, i.e. to acquire knowledge on the real part by measurements of

the imaginary part over the whole spectrum or vice versa.

Nevertheless, in spite of such important formal analogies, it is important to stress some

qualitative differences between equilibrium and non-equilibrium systems in the physical

meaning of the linear response function. Whereas in systems close to equilibrium there

is basically equivalence between the natural fluctuations and the linear response to external

perturbations, as clarified by the fluctuation-dissipation theorem [11, 12], in the considered

non equilibrium systems such symmetry in broken, the mathematical reason being that the

SRB measure is smooth only along the unstable manifold. A more geometrical view of this

fact is that, whereas natural fluctuations of the system are restricted to the unstable mani-

fold, because, by definition, asymptotically there is no dynamics along the stable manifold,

external forcings will cause almost always motions having components - of exponentially

decaying amplitude - out of the unstable manifold [3, 4]. For a discussion of this point, see

also [22]. It should also be noted that the non-equivalence of forced and free fluctuations in

chaotic systems was already pointed out and tackled in heuristic terms in the late ’70s by

Lorenz [13] when considering the atmospheric system.

Whereas K-K relations for linear processes, thanks to their generality, have become a

basic textbook subject and standard tool in many different fields, such as acoustics, signal

processing, optics, statistical mechanics, condensed matter physics, material science, rela-

tively little attention has been paid to theoretical and experimental investigation of K-K

relations and sum rules of the nonlinear susceptibilities, in spite of the ever increasing scien-

tific and technological relevance of nonlinear physical processes. Recently, several theoretical

and experimental results in this direction have been formulated in the context of analyzing

nonlinear processes of interaction of radiation with matter [14, 15].

The main goal of this paper is to analyze the formal properties on the nth order pertur-

bative response of Axiom A, non equilibrium steady state systems to external forcings. In

particular, we develop a theory of generalized K-K relations that extend, on one side, the

results on the linear case given by Ruelle [3] for this class of systems, and on the other side,

what obtained for nonlinear processes in electronic systems close to equilibrium [14, 15] and

in simple yet prototypical mechanical systems [18]. Special attention is paid to the case

of nonlinear susceptibilities describing processes responsible for harmonic generation, whose

properties are such that a rather extensive set of important properties - including sum rules

- can be deduced. We stress that also in the nonlinear case K-K relations constitute un-

avoidable benchmarks that any experimental and model generated dataset must obey. K-K

relations may prove, as discussed later, useful tools for defining a theory of climate change,

because they apply also for systems where the fluctuation-dissipation theorem is not verified.

Our paper is structured as follows. In Sec. II, we introduce the properties of the gen-

eral nth susceptibility, resulting as Fourier transform of the nth perturbative order response

function of the system. In Sec. III, we present an extension of the theory of nonlinear K-K

relations to the dynamical systems considered by Ruelle, showing which class of nonlinear

susceptibilities obey K-K relations and deducing rigorous results in the case of harmonic

generation processes. In Sec. IV, we discuss our results, present our conclusions and per-

spectives for future investigations. Two appendices are also included. In App. A we show

how the Kubo theory can be formally reconciled with the results by Ruelle, so that the

results presented in this work can be applied also for general equilibrium systems. A discus-

sion of the relevance for climate studies of the response theory for Axiom-A systems and of

the specific results described in this study is given in App. B.

II. LINEAR AND NONLINEAR RESPONSE OF PERTURBED NON-

EQUILIBRIUM STEADY STATES

We consider an autonomous Axiom-A flow x = F (x) defined in a compact manifold,

such that x(t) = f tx, with x = x(0). The flow is assumed to possess a chaotic, mixing

dynamic, and to be associated to an invariant ergodic SRB measure ρSRB(dx), such that for

any measurable observable Φ(x) the ensemble average is equal to the time average:

〈Φ〉0 =

ρSRB(dx)Φ(x) = limT→∞

1

T

T∫

0

dtΦ(f tx) = limT→∞

dxΦ(fT x) (1)

for almost every initial condition x according to the Lebesgue measure dx; the last equality

holds for the special case of mixing dynamics. The SRB measure is usually singular, but

smooth along the directions of the unstable manifold [5, 6, 19]. Ruelle has shown that

for this specific class of dynamical systems (as well as for the corresponding discrete-time

diffeomorphisms) it is possible to differentiate the SRB states [20, 21] when the flow is

perturbed by an infinitesimal vector field in the following way:

x = F (x) + e(t)X(x). (2)

Is is then possible to express the perturbed expectation value of Φ(x) in terms of a pertur-

bation series:

〈Φ〉(t) = 〈Φ〉0 +

∞∑

n=1

〈Φ〉(n)(t) (3)

where, proposing a slight generalization of the formula proposed by Ruelle [4], which consid-

ered purely periodic perturbations, the nth term can be expressed as a n-uple convolution

integral of the nth order Green function with n terms each representing the suitably delayed

time modulation of the perturbative vector field:

〈Φ〉(n)(t) =

∞∫

−∞

∞∫

−∞

. . .

∞∫

−∞

dσ1dσ2 . . .dσnG(n)(σ1, . . . , σn)e(t − σ1)e(t − σ2) . . . e(t − σn). (4)

The nth order Green function is causal, i.e. its value is zero if any of the argument is

non positive, and can be expressed as time dependent expectation value of an observable

evaluated over the unperturbed SRB measure:

G(n)(σ1, . . . , σn) =

ρSRB(dx) Θ(σ1)Θ(σ2 − σ1) . . .Θ(σn − σn−1) ×

×ΛΠ(σn − σn−1) . . .ΛΠ(σ2 − σ1)ΛΠ(σ1)Φ(x), (5)

where Θ is the usual Heaviside function, Λ(•) = X(x)∇(•) describes the effect of the

perturbative vector field, and Π induces the time evolution along the unperturbed vector

field so that Π(τ)A(x) = A(x(τ)) for any observable A. The n = 1 term describes the

linear response of the system to the perturbation field [3], and, thanks to the superposition

principle, can be derived also by using the method of impulse perturbation [22]. In App.

A we show that it is possible to rephrase the Kubo response theory [1] in such a way to

obtain a formula that perfectly matches the formula presented in Eq. (5), provided that the

equilibrium canonical distribution is used instead of the SRB measure ρSRB(dx).

A. Response of the system in the frequency domain

If we compute the Fourier transform of the nth order perturbation to the expectation

value 〈Φ〉n(t) defined in Eq. (4) we obtain:

〈Φ〉(n)(ω) =

∞∫

−∞

. . .

∞∫

−∞

dω1 . . . dωnχ(n) (ω1, . . . , ωn) e(ω1) . . . e(ωn) × δ

(

ω −

n∑

l=1

ωl

)

, (6)

where the Dirac δ guarantees that the sum of the arguments of the Fourier transforms of the

time modulation functions equals the argument of the Fourier transform of 〈Φ〉n(t), whereas

the susceptibility function is defined as

χ(n) (ω1, . . . , ωn) =

∞∫

−∞

. . .

∞∫

−∞

dt1 . . .dtnG(n) (t1, . . . , tn) exp

[

in∑

j=1

ωjtj

]

. (7)

These operations make sense if the Green function is integrable or at least, in a weaker,

distributional sense, if it not exponentially increasing. In the linear case, Ruelle [3, 4]

has shown that integrability is ensured by proving that both the contributions associated

to terms resulting from projections of the perturbative vector field on the unstable and

stable manifolds converge, because of the distinct processes of mixing and of exponential

contraction, respectively. In the nonlinear n > 1 case, we can heuristically use the same

arguments - as well as taking into account that in the classical equilibrium case [1] the

higher order correlations are typically much weaker and with faster decrease - to exclude

the possibility that the operation presented in Eq. (7) is meaningless.

Assuming that, without serious loss of generality, the function e(t) can be expressed as:

e(t) =m∑

k=1

eωjexp[−iωjt] + e−ωj

exp[iωjt] (8)

with eωj=[

e−ωj

]∗, we derive that each frequency component in Eq. 6 can be written as:

〈Φ〉(n)(ω) =∑

ωΣ

〈Φ〉(n)

(ωΣ) δ (ω − ωΣ) , (9)

where we are summing over all the possible distinct values ωΣ of the possible sums of n

among the 2m frequencies in the spectrum of e(t), which basically formalizes the process of

frequency mixing. Of course, in the linear n = 1 case, no mixing occurs and outputs can be

observed at the same frequencies as the input. In general, each term 〈Φ〉(n)

(ωΣ) is given by

the following sum:

〈Φ〉(n)

(ωΣ) =∑

P

ωkj=ωΣ

χ(n) (ωk1, . . . , ωkn

) ewk1. . . ewkn

, (10)

where the sum of the arguments of all the contributing susceptibility functions is ωΣ. Note

that, from an experimental point of view, we can measure 〈Φ〉(n)

(ωΣ) by analyzing in the

frequency domain the perturbed output of the system, whereas disentangling the various

terms contributing to the summation in Eq. (10) is rather hard. Again, this problem is not

present in the linear case.

III. GENERALIZED KRAMERS-KRONIG RELATIONS

A. Basic Results

Once we are granted that at every order n the response on the system 〈Φ〉(n)(t) is written

as a convolution integral having as Kernel a causal Green function G(n)(σ1, . . . , σn), and

assuming that the suitable integrability conditions are obeyed, we are in the conditions of

writing generalized dispersion relations for the nth order susceptibility presented in Eq. (7),

along the lines of what developed in the context of optics in [14, 15]. Therefore, we can

apply Titchmarsch’s theorem [9, 10, 14, 15] separately to each variable of G(n)(σ1, . . . , σn)

and deduce that χ(n) (ω1, . . . , ωn) is holomorphic in the upper complex plane of each variable

ωi, 1 ≤ i ≤ n. If we consider the first argument ω1 of the nonlinear susceptibility function

(7), the following dispersion relation holds

P

∞∫

−∞

dω′1

χ(n) (ω′1, . . . , ω

′n)

ω′1 − ω1

= iπχ (ω1, . . . , ω′n) , (11)

where P indicates that integration is performed in principal part. Repeating the same

procedure for all the remaining n − 1 frequency variables, we obtain

P

∞∫

−∞

. . .

∞∫

−∞

dω′1 . . .dω′

n

χ(n) (ω′1, . . . , ω

′n)

(ω′1 − ω1) . . . (ω′

n − ωn)= (iπ)n χ (ω1, . . . , ωn) , (12)

which extends to all orders the linear K-K relations already described by Ruelle [3]. K-K

relations constitute self-consistency constraints that must be obeyed and allow to reconstruct

the real part of the response from the imaginary part, or vice-versa. The principle of causality

of the response function is reflected mathematically in the validity of the K-K relations

presented in Eq. (12). Note that, as discussed by Peiponen [16, 17], all the functions[

χ(n) (ω1, . . . , ωn)]m

, with m ≥ 1 obey the very same dispersion relations as that written for

m = 1 in Eq. (12). This implies that the generality of these dispersion relation goes beyond

not only the distinction between classical and quantum equilibrium system, as discussed in

[14, 15], but also beyond the distinction between equilibrium and non-equilibrium systems,

at least when we consider the Axiom A case or adopt the chaotic hypothesis for many

particles systems.

The dispersion relations (11) and (12) may be thought of being of doubtful interest from

an experimental point of view, since on one side we basically can have access to quantities like

〈Φ〉(n)

(ωΣ), which results from a linear combination of, in general, more than one different

susceptibility functions (in the sense that they are evaluated at different values of their

arguments). Moreover, most of the physically relevant nonlinear phenomena are described by

nonlinear susceptibilities where all or part of the frequency variables are mutually dependent,

such in the later described case of nth order harmonic generation at frequency nω in the

presence of a monochromatic modulation function e(t) = exp [−iω0t]+exp [iω0t] of frequency

ω0. We may, therefore, understand that a more flexible theory is needed in order to provide

the effectively relevant dispersion relations for nonlinear phenomena.

B. A New Definition of Dispersion Relations

We take the following point of view. When considering the nth order nonlinear process, a

meaningful dispersion relation involves a line integral in the space of the frequency variables,

which entails the choice of a one-dimensional space embedded in a n-dimensional space.

This corresponds to the realistic experimental setting where only the frequency of one of

the monochromatic fields described in Eq. (8) is changed. Since in the nonlinear setting we

have frequency mixing, changing the frequency of one of the components of the forcing will

change differently each of the terms 〈Φ〉(n)

(ωΣ), depending on whether none, one or more

than one arguments of the contributing nonlinear susceptibility functions (see Eq. (10))

are varied. The choice of the parameterization then selects different susceptibilities and so

refers to different nonlinear processes. Each component j of the straight line in Rn can be

parameterized as follows:

ωj = vjs + wj, 1 ≤ j ≤ n, (13)

where the parameter s ∈ R, the vector ~v ∈ Rn of its coefficients describes the direction of the

straight line, and the vector ~w ∈ Rn determines ~ω(0). Since we know that χ(n) (ω1, . . . , ωn)

is holomorphic in the upper complex plane of each variable ωi, 1 ≤ i ≤ n, we have that the

extension for complex values of s of the function:

χ(n) (s) = χ(n) (v1s + w1, . . . , vns + wn)

=

∞∫

−∞

. . .

∞∫

−∞

dt1 . . .dtnG(n) (t1, . . . , tn) exp

[

isn∑

j=1

vjtj + in∑

j=1

wjtj

]

(14)

is holomorphic in the upper complex s plane if all the components of vector ~v are non-

negative. This construction has been first proposed in the context of nonlinear optics in [23].

Hence, by applying the Titchmarch theorem, we deduce that for all m ≥ 1 the following

integral relation holds for the susceptibility defined in Eq. (14):

iπ[

χ(n)(s)]m

= P

∞∫

−∞

[

χ(n)(s′)]m

s′ − sds′, (15)

which, when the real and imaginary part of the nonlinear susceptibility are considered,

results into:

Re

[

χ(n)(s)]m

=1

πP

∞∫

−∞

Im[

χ(n)(s′)]m

s′ − sds′, (16)

Im

[

χ(n)(s)]m

= −1

πP

∞∫

−∞

Re[

χ(n)(s′)]m

s′ − sds′. (17)

The condition on the sign of the directional vectors of the straight line in Rn implies that only

one particular class of nonlinear susceptibilities possess the holomorphic properties required

to obey the dispersion relations (16). Hence, causality is not a sufficient condition for

the existence of K-K relations between the real and imaginary part of a general nonlinear

susceptibility function, if its arguments are mutually dependent. We stress that, instead,

causality implies that Eq. (12) holds.

C. Harmonic Generation

In order to clarify the results presented in the previous sections, and show how they

can be used for analyzing actual data, we concentrate on the simplified setting of a single

monochromatic perturbation field such that e(t) = exp[−iω0t] + exp[iω0t]. In this case, at

each order n, ωΣ = ±(2j + 1), with ω0, j = 0, . . . , (n − 1)/2 if n is odd and ωΣ = ±2jω0,

with j = 0, . . . , n/2 if n is even. Note that for even orders there is always a static response,

which, in the optical literature, is known as optical rectification [15]. If we focus, e.g., on

the third order of perturbation and consider only the positive frequencies, we have that the

observable signal at ω0, constituting the first correction to the linear response is:

〈φ〉(3)

(ω0) = χ(3)(−ω0, ω0, ω0) + χ(3)(ω0,−ω0, ω0) + χ(3)(ω0, ω0,−ω0); (18)

whereas the observable signal, responsible for the third harmonic generation is:

〈φ〉(3)

(3ω0) = χ(3)(ω0, ω0, ω0); (19)

If we change the frequency ω0 of the perturbation field and study how the output varies, it

is clear that for all the three terms comparing on the right hand side of Eq. (18) the vector

~v of the s-parameterization proposed in Eq. (13) has one negative component, whereas

~v = (1, 1, 1) for the only term responsible for harmonic generation in Eq. (19). This implies

that, when analyzing the first nonlinear correction to the linear response at frequency ω0,

we cannot expect that K-K relations apply, since poles in the upper complex plane of the

s = ω0 may well be present [24]. In this case, different signal processing techniques, such as

the Maximum Entropy Method, have to be adopted [15]. Therefore, the condition on the

sign of ~v is in this case useful for giving a negative statement, i.e. determining when K-K

cannot be applied. On the contrary, we are granted that the susceptibility describing the

third harmonic nonlinear response obeys K-K relation. It is clear that the same applies at

all orders n, and also it can be easily shown that the only contribution to the observable

〈φ〉(n)

(nω0) is χ(n)(ω0, . . . , ω0).

At every order n, we have that χ(n) (−ω1, . . . ,−ωn) =

χ(n) (ω1, . . . , ωn)∗

(Z∗ indicat-

ing the complex conjugate of Z), because 〈Φ〉(n)(t) and e(t) are real. It is is easy to show

that the following relation holds for all values of m ≥ 1:

[

χ(n) (−ω1, . . . ,−ωn)]m

=

[

χ(n) (ω1, . . . , ωn)]m∗

, (20)

We then derive that at all orders n ≥ 1 and for all m ≥ 1:

−π

2Im

[

χ(n) (ω0, . . . , ω0)]m

= ω0P

∞∫

0

dω′0

Re[

χ(n) (ω′0, . . . , ω

′0)]m

(ω′02 − ω2

0)(21)

π

2Re

[

χ(n) (ω0, . . . , ω0)]m

= P

∞∫

0

dω′0

ω′0Im

[

χ(n) (ω′0, . . . , ω

′0)]m

(ω′02 − ω2

0)(22)

which, albeit in a different perspective from what shown in Eq. (12), generalize the linear

K-K at all orders. Note that, if we consider limω0→0 of Eq. (22) in the linear case and assume

that the limits converge, we obtain the following expression for the linear static response of

the system:

Re

[

χ(1) (0)]m

=[

Re

χ(1) (0)]m

=2

πP

∞∫

0

dω′0

Im[

χ(1) (ω′0)]m

ω′0

; (23)

the finiteness of the integral is consistent with the fact that, by symmetry,

Im[

χ(1) (ω0 = 0)]m

= 0, which must be obeyed for all values of m ≥ 1. Note that a

detailed verification of linear K-K has been performed in the case of Lorenz system [25].

It is somewhat surprising to observe how the qualitative features of the detected (and re-

constructed) susceptibility are similar to what results from a simple oscillator model: the

imaginary part has a strong peak for a resonance of system (even if in this case there is

no deterministic natural frequency for the system), which matches the dispersive structure

found for the real part of the susceptibility. Another minor spectral feature is observed, and

again, following the spectroscopic paradigm, a peak in the imaginary part is associated to

a dispersive structure in the real part. Note also that the Lorenz system is non-Axiom A,

which suggests that a wide range of applicability for these relations is still to be explored.

Note that, even if several monochromatic forcings are present, Eqs. (21)-(22) still apply,

since no other frequency components are involved.

If we plug ~v = (1, . . . , 1) and ~w = (0, . . . , 0), and redefine s = ω0 in Eq. (14), and consider

the basic properties of the Fourier Transform, we have that the short time behavior of the

nth order Green function determines the asymptotic behavior of the nth order harmonic

susceptibility at frequency nω0. We perform the following variable change

tj =

j∑

k=1

τk, (24)

assume that G(n)(t1(τ1), . . . , tn(τ1, . . . τn)) be smooth for all its arguments τj in 0, and let β

be the smallest sum of exponents of (τ1, . . . τn) such that there is a non-vanishing monomial

Mβ(τ1, . . . , τn) in the Taylor expansion G(n)(t1(τ1), . . . , tn(τ1, . . . τn)). We then have that the

following limit exists and is finite [14, 15, 26]:

limω0→∞

ωβ+n0 χ(n)(ω0, . . . , ω0) = α ∈ R \ 0, (25)

which implies that the asymptotic behavior of χ(n)(ω0, . . . , ω0) is at least as fast as ω−n0 .

Moreover, since Re

χ(n)(ω0, . . . , ω0)

is an even function of ω0 and, from Eq. (25), deter-

mines the asymptotic behaviour (Im

χ(n)(ω0, . . . , ω0)

has a faster asymptotic decrease),

we derive that β + n must be even, so that β + n = 2γ. Therefore, dispersion theory

provides us with indirect information also about the short time behavior of the Green func-

tion. Furthermore, the knowledge of the asymptotic behavior allows a further generaliza-

tion of what presented in (21)-(22). In fact, we have that all the (independent) functions

ω2pχ(n)(ω0, . . . , ω0), p = 0, . . . , γ − 1 are holomorphic in the upper complex plane of ω0 and

obey suitable integrability conditions allowing for writing the following set of generalized

K-K relations:

−π

2ω0

2p+1Im

[

χ(n) (ω0, . . . , ω0)]m

= P

∞∫

0

dω′0

ω′02pRe

[

χ(n) (ω′0, . . . , ω

′0)]m

((ω′02 − ω2

0), (26)

π

2ω0

2pRe

[

χ(n) (ω0, . . . , ω0)]m

= P

∞∫

0

dω′0

ω′02p+1Im

[

χ(n) (ω′0, . . . , ω

′0)]m

(ω′02 − ω2

0). (27)

with p = 0, . . . , mγ − 1. Comparing the asymptotic behavior given in Eq. (25) with those

obtained by applying the superconvergence theorem [27] to the general K-K relations (26)-

(27), we derive the following set of general sum rules

∞∫

0

ω0′2pRe

[

χ(n) (ω′0, . . . , ω

′0)]m

dω′ = 0, 0 ≤ p ≤ mγ − 1, (28)

∞∫

0

ω0′2p+1Im

[

χ(n) (ω′0, . . . , ω

′0)]m

dω′ = 0, 0 ≤ p ≤ mγ − 2, (29)

∞∫

0

ω0′2p+1Im

[

χ(n) (ω′0, . . . , ω

′0)]m

dω′ = −αmπ

2, p = mγ − 1. (30)

All the moments of the nth order harmonic generation susceptibility vanish except that of

order 2γ−1 of the imaginary part. This latter sum rule creates a conceptual bridge between

the measurements of the imaginary part of the susceptibility under examination throughout

the spectrum to the short term behavior of the nth Green function. These results hold for

all values of m ≥ 1. The generalized K-K relations and sum rules here presented constitute

a rather extensive set of stringent integral constraints that must be obeyed by experimental

data and model simulations. These results generalize what obtained for general optical

systems near equilibrium [26] and for simple mechanical systems [18]. Note that both the

generalized K-K relations (26)-(27) and the sum rules (28)-(30) have been verified in detail

on experimental data in the case of optical processes near equilibrium [28].

IV. SUMMARY AND CONCLUSIONS

In this paper we have considered the general response function G(n)(t1, . . . , tn) recently

proposed by Ruelle [3, 4] for describing the impact of small time-dependent forcings to

the non-equilibrium steady states resulting from Axiom A dynamical systems, which, when

taking into account the chaotic hypothesis by Gallavotti and Cohen [7, 8], are of general

physical interest. At all orders of perturbative expansion, the effect of the forcing on the

expectation value of a general observable can be expressed in terms of means of quantities

performed at non-equilibrium steady state.

Since, at every order of perturbation, the response function is causal, it is possible to

write a set of Kramers-Kronig relations for the corresponding susceptibility, defined as the

multivariable Fourier Transform of the response function χ(n)(ω1, . . . , ωn). These dispersion

relations are of little applicability because they cannot be used to effectively analyze the

output signal, which is the change in the expectation observable of the considered observable.

In practice, it is interesting to consider the case of one or more monochromatic forcings

and to be in the condition of analyzing what happens when the frequency of one of them is

changed. Since in the nonlinear setting of order n we have frequency mixing, such frequency

tuning will affect differently the various frequency components of the observed output signal,

depending on whether none, one or more than one arguments of the nonlinear susceptibil-

ity functions responsible for the observed frequency components of the output are varied.

Therefore, following this approach, the dispersion relation becomes a parameterized line in-

tegral in the n-dimensional space of frequency variables. K-K relations apply only for special

forms of parameterizations, which correspond to a specific family of susceptibility functions.

These results are system-independent and derive strictly from complex analysis.

Among the phenomena which can be treated using the K-K formalism, we concentrate

on the nth order process by which the system responds at frequency nω0 when forced by a

monochromatic vectorial field with angular frequency ω0. Such a process is described by the

harmonic generation susceptibility χ(n)(ω0, . . . , ω0), which is holomorphic in the upper com-

plex ω0 plane and obeys K-K relations. For any given system, the asymptotic behavior for

large frequencies is shown to depend on the short-time response and to be of the form ω−2γ0 .

It is then proved that all functions ω2pχ(n)(ω0, . . . , ω0) with p = 0, . . . , γ − 1 obey K-K re-

lations, so that more stringent , generalized constraints are established. Furthermore, using

symmetry arguments and the superconvergence theorem on the generalized K-K relations,

and comparing the results with the asymptotic behavior for large values of ω0, new sum rules

are obtained. We derive that all even moments of the real part and all odd moments of the

imaginary parts are null, except for the highest converging odd moment of the imaginary

part of the susceptibility, which is directly related to the short time behavior of the system.

Furthermore, these results are also extended to the powers[

χ(n)(ω0, . . . , ω0)]m

, m ≥ 1 of

the susceptibility, and additional constraints are derived. The obtained generalized K-K

relations and sum rules can be used to check any experimental data and approximate theory

of nonlinear phenomena, because they are necessary constraints which have to be obeyed.

These results generalize and extend what obtained by Ruelle [3, 4] for Axiom A systems,

set in a much more general theoretical framework previous findings obtained for near equi-

librium optical processes [14, 15] and and simple yet prototypical mechanical systems near

equilibrium [18], and shed light on the generality of the constraints deriving from the prin-

ciple of causality, which can be used for testing model outputs and experimental data, both

for equilibrium and non-equilibrium systems. Note that, as discussed in [10, 14, 15], basi-

cally all K-K relations and sum rules can be rephrased, after lengthy but straightforward

calculations, in terms of absolute value and phase of the susceptibility function, which in

some cases may be of easier experimental observation.

It is somewhat surprising, and encouraging in the perspective of the theory here devel-

oped, to see that the linear susceptibility of the Lorenz system investigated in [25] looks a lot

like the result of an optical experiment: the peaks of the imaginary part, corresponding to

the resonances of system (even if in this case there are no deterministic natural frequencies)

match the dispersive structures found for the real part of the susceptibility. Note that, far

from being a curiosity, it is through this approach that the optical constants of most solids

have been actually computed [29, 30].

In order to clarify and complete the picture, we have shown, in App. A, that the functions

derived for non equilibrium steady states are formally equivalent, at all perturbative orders,

to what obtained with the Kubo formalism for the response of systems close to equilibrium,

apart from the measure involved in the phase space integration. In the case of near to

equilibrium system, the measure is the one describing the canonical distribution, whereas

in the setting analyzed by Ruelle, the SRB measure of the unperturbed flow is involved.

Therefore, all the results presented in the paper apply, a fortiori, for these equilibrium

systems.

The response theory for Axiom A systems can have interesting implications for climate

studies. In fact, the possibility of defining a response function basically poses the problem of

climate change is well-defined context, and, when considering the procedures aimed at im-

proving climate models, justifies rigorously the procedures of tuning and adjusting of the free

parameters. Furthermore, qualitative differences between different and widespread ensemble

simulation practices can be interpreted in this context. Moreover, the non-equivalence of free

and forced fluctuations explains why many attempts of applying the fluctuation-dissipation

theorem in climate studies have basically failed. Instead, it may be that the general theory

of Kramers-Kronig relations described in this paper, which, in the case of non-equilibrium

system, is decoupled form the fluctuation-dissipation theorem, may provide a viable way of

defining a comprehensive self-consistent theory of climate change, ensured by the integral

relations connecting the in-phase and our of phase components of the response of the system

to external perturbations. This is discussed in some greater detail in App. B.

We conclude with some practical caveats. As well known, it is surely not trivial in

practical terms to effectively verify the K-K relations and sum rules on experimental or model

generated data. One general problem is their integral formulation, which requires that data

are available on a rather extensive spectral range and with a reasonable resolution. This may

raise issue of computational costs and/or experimental set-up. The extrapolations in K-K

analysis can be a serious source of errors [30, 31]. Recently, King [32] presented an efficient

numerical approach to the evaluation of K-K relations, and singly and multiply subtractive

K-K relations have been proposed in order to relax the limitations caused by finite-range data

[33, 34]. It should be noted that K-K relations for higher-order susceptibilities are, somewhat

counter-intuitively, sometimes easier to verify than the linear K-K relations, because they

have typically a much faster asymptotic decrease. Whereas we have shown that at all

orders large families of K-K relations hold for the various moments and various powers of

the susceptibility functions, it should be expected that they do not converge at the same

rate when data of finite precisions coming from a finite spectral range are used. See the

discussion in [14, 15]. Furthermore, when considering chaotic systems, further problems in

signal detection of the system response at specific frequencies are related to the presence

of a continuous spectrum in the background; this latter issue may become more serious

when nonlinear processes are examined and the observed monochromatic signal is weaker.

Nevertheless, along the line of Reick [25] these problems may result to be manageable.

Acknowledgments

The author wishes to thank F. Bassani, K.-E. Peiponen, D. Ruelle, and A. Speranza for

crucial intellectual stimulations.

APPENDIX A: RECONCILING KUBO’S AND RUELLE’S GENERAL PERTUR-

BATIVE RESPONSE FUNCTIONS

In this appendix we show how to reconcile formally the nth order perturbative response

for general systems characterized by non-equilibrium steady state presented in Eq. (5) with

the classical results obtained with the Kubo formalism [1] for the response for systems close

to equilibrium. Therefore, all the results presented in the paper apply, a fortiori, for these

equilibrium systems.

We consider a system of N degrees of freedom described by the canonical coordinates

q = (q1, . . . , qN) and p = (p1, . . . , pN) and evolving under the action of the Hamiltonian

operator H(q, p) = H0(q, p) + h(q, p, t), composed of the unperturbed Hamiltonian H0(q, p)

plus the time dependent perturbation (weak) Hamiltonian expressed in the form h(q, p, t) =

−e(t)B(q, p) [2]. The evolution equation of the system can then be written as:

x = F (x) + e(t)X(x) (A1)

where x = (q, p); F (x) = Ω∇H(x), X(x) = −Ω∇B(x), with Ω indicating the symplectic

matrix. We assume that, if the perturbation is set to 0, the expectation value of any

observable Φ can be expressed as the following:

〈Φ〉0 =

dxρ0(x)Φ(x) =

ρ0(dx)Φ(x) (A2)

where integration is performed in the phase space of the system, and the canonical distri-

bution, which is absolutely continuous with respect to the Lebesgue measure of the phase

space, is defined as usual as:

ρ0(dx) = ρ0(x)dx =exp [−H0(x)/kT ]

dΓ exp [−H0(x)/kT ]dx =

exp [−H0(x)/kT ]∫

ρ0(dx)dx. (A3)

Following the perturbative approach introduced by Kubo [1], we have that, for small per-

turbations, the expectation value of Φ at time t can be written as:

〈Φ〉(t) = 〈Φ〉0 +∞∑

n=1

〈Φ〉(n)(t) (A4)

where the terms under summation describe the non-equilibrium properties - for a system

which is close to equilibrium - at all orders of perturbation; in particular the n = 1 terms

provides information of the linear response of the system. The perturbative terms can be

expressed as follows [1, 2]:

〈Φ〉(n)(t) =

∞∫

−∞

∞∫

−∞

. . .

∞∫

−∞

dσ1dσ2 . . .dσn ×

×Θ(σ1)Θ(σ2 − σ1) . . .Θ(σn − σn−1)f(t − σ1)f(t − σ2) . . . f(t − σn) ×

×〈[B(x), . . . [B(x(σn − σ2)), . . . [B(x(σn − σ1)), Φ(x(σn))]] . . .]〉0 =

=

∞∫

−∞

∞∫

−∞

. . .

∞∫

−∞

dσ1dσ2 . . .dσnG(n)(σ1, . . . , σn)e(t − σ1)e(t − σ2) . . . e(t − σn), (A5)

where the time evolution of the observables B and Φ is due to the unperturbed Hamiltonian

H0. The two main are that the non-equilibrium response is expressed as a convolution

integral where the Kernel, which is the nth order Green function G(n)(σ1, . . . , σn) is causal.

Note that the Kernel operator in the quantum case, where we deal with N-particles

Hilbert space and observables are replaced by operators, is formally obtained by simply

substituting each Poisson brackets [•, •] with 1/(i~) times the commutator •, •, and by

redefining the expectation value at equilibrium of a generic operator P as follows:

〈P 〉0 =

a〈a|P |a〉 exp [−Ea/kT ]∑

b exp [−Eb/kT ]. (A6)

where |a〉 is the eigenstate with eigenvalue Ea of the Hamiltonian operator H0.

Since the following trivial identity holds:

[B(x), •] = X(x)∇(•) = Λ(•) (A7)

and since, by definition, the evolution of any observable A driven by the unperturbed Hamil-

tonian H0 can be formally represented as follows:

A(x(τ)) = exp(iτL)A(x) = Π(τ)A(x), (A8)

where iLA(x) = [A(x), H0(x)], the nth order Green function can be formally written in the

following compact and form:

G(n)(σ1, . . . , σn) =

ρ0(dx)ΛΠ(σn − σn−1) . . .ΛΠ(σ2 − σ1)ΛΠ(σ1)Φ(x). (A9)

which is fully equivalent to the formula shown in Eq. (5), provided that the measure de-

scribing the equilibrium canonical distribution is substituted with the general SRB measure.

APPENDIX B: RESPONSE THEORY FOR NON-EQUILIBRIUM STEADY

STATES AND CLIMATE RESEARCH

When adopting the chaotic hypothesis, the possibility of defining a response function of

a perturbed non-equilibrium steady state and its actual properties seem to have very inter-

esting impacts in climate studies. On one side, this creates a context where the problem of

climate change is well-posed at mathematical level and where, when considering the proce-

dures aimed at improving climate models, the tuning and adjustment of the free parameters

- at least locally - may be considered as a well-defined operation devoid of catastrophic

impacts on the statistical properties of the system. On the other hand, straightforward ap-

plications of fluctuation-dissipation theorem [35, 36], or the idea that climate change signals

project on the natural modes of climate variability [37] seem inadequate, as discussed in [38].

Instead, it seems that the theory of Kramers-Kronig relations described in this paper may

provide a viable way of defining a comprehensive self-consistent theory of climate change,

ensured by the integral relations connecting the in-phase and our of phase components of

the response of the system to external perturbations. As an example, we may interpret

Eq. (23) as the fact that the static response function - measuring climate sensitivity - can

be related to the out-of-phase response to same forcing at all frequencies, at least in first

approximation.

The concepts behind the Ruelle response theory also clarify the meaning of some com-

mon ensemble simulation practices, which are widely adopted by the climate modelling

community with the goal of estimating the uncertainty on the statistical properties of the

model outputs, when a specific set of observables is considered [39, 40, 41]. Three different

strategies, which are nevertheless more and more hybridized, can be pointed out:

• Each simulation is performed with the same climate model, but starting from slightly

different initial state;

• Each simulation is performed with the same climate model, but with slightly different

values of some key uncertain parameters characterizing the global climatic properties,

• Each simulation is performed with a different climate model (multi-model ensemble).

Under the chaotic hypothesis, the first procedure seems useful, since a more detailed explo-

ration of the phase space of the system, with a better sampling - on a finite time - of the

attractor of the model. The significance of the second procedure seem to be reinforced by the

response theory for non equilibrium steady states, because in this case the variously tuned

models basically explore parameterically deformed ergodic measures, and the macroscopic

sensitivity of the model is thus explored. As for the third procedure, whereas it surely al-

lows for climate model intercomparison, aggregating information from from rather different

attractors seems ill-defined.

[1] R. Kubo: Statistical-mechanical theory of irreversible processes. I, J. Phys. Soc. Jpn. 12

(1957), 570-586.

[2] D. N. Zubarev: Nonequilibrium Statistical Thermodynamics, Consultant Bureau, New York,

1974.

[3] D. Ruelle: General linear response formula in statistical mechanics, and the fluctuation-

dissipation theorem far from equilibrium. Phys. Letters A 245 (1998), 220–224.

[4] D. Ruelle: Nonequilibrium statistical mechanics near equilibrium: computing higher order

terms. Nonlinearity 11 (1998), 5–18.

[5] J.-P. Eckmann, D. Ruelle: Ergodic theory of chaos and strange attractors, Rev. Mod. Phys.

57 (1985), 617–655.

[6] D. Ruelle: Chaotic evolution and strange attractors, Cambridge University Press, Cambridge,

1989.

[7] G. Gallavotti and E.G.D. Cohen: Dynamical ensembles in stationary states, J. Stat. Phys. 80

(1995), 931–970.

[8] G. Gallavotti: Chaotic hypothesis: Onsager reciprocity and fluctuation-dissipation theorem,

J. Stat. Phys. 84 (1996), 899–926.

[9] H. M. Nussenzveig: Causality and Dispersion Relations, Academic, New York, 1972.

[10] K.-E. Peiponen, E. M. Vartiainen, and T. Asakura: Dispersion, Complex Analysis and Optical

Spectroscopy, Springer, Heidelberg, 1999.

[11] J. Weber: Fluctuation Dissipation Theorem, Phys. Rev. 101 (1956), 1620–1626.

[12] R. Kubo: The Fluctuation Dissipation Theorem, Rep. Prog. Phys. 29 (1966), 255–284.

[13] E.N. Lorenz: Forced and Free Variations of Weather and Climate, J. Atmos. Sci. 36 (1979),

1367–1376.

[14] V. Lucarini, F. Bassani, J.J. Saarinen, and K.-E. Peiponen: Dispersion theory and sum rules

in linear and nonlinear optics, Rivista del Nuovo Cimento 26 (2003), 1-120.

[15] V. Lucarini, J.J. Saarinen, K.-E. Peiponen, E. Vartiainen: Kramers-Kronig Relations in Op-

tical Materials Research, Springer, Heidelberg, 2005.

[16] K.-E. Peiponen: Sum rules for the nonlinear susceptibilities in the case of sum frequency

generation, Phys. Rev. B 35 (1987), 4116–4117.

[17] K.-E. Peiponen: Nonlinear susceptibilities as a function of several complex angular-frequency

variables, Phys. Rev. B 37 (1988), 6463–6467.

[18] F. Bassani and V. Lucarini: General properties of optical harmonic generation from a simple

oscillator model, Il Nuovo Cimento D 20 (1998), 1117–1125.

[19] L.S. Young: What are SRB measures, and which dynamical systems have them?, J. Stat.

Phys. 108(5) (2002), 733-754.

[20] D. Ruelle: Differentiation of SRB states. Commun. Math. Phys. 187 (1997), 227–241.

[21] D. Ruelle: Differentiation of SRB states: correction and complements. Commun. Math. Phys.

234 (2003), 185-190.

[22] B. Cessac, J.-A. Sepulchre: Linear response, susceptibility and resonances in chaotic toy

models, Physica D 225 (2007), 13–28.

[23] F. Bassani and S. Scandolo: Dispersion relations and sum rules in nonlinear optics, Phys. Rev.

B 44 (1991), 8446–8453.

[24] K.-E. Peiponen, J. J. Saarinen, and Y. Svirko: Derivation of general dispersion relations and

sum rules for meromorphic nonlinear optical spectroscopy, Phys. Rev. A 69 (2004), 043818.

[25] C.H. Reick: Linear response of the Lorenz system, Phys. Rev. E 66 (2002), 036103.

[26] F. Bassani and V. Lucarini: Asymptotic behaviour and general properties of harmonic gener-

ation susceptibilities, Eur. Phys. J. B 17 (2000), 567–573.

[27] G. Frye and R. L. Warnock: Analysis of partial-wave dispersion relations, Phys. Rev. 130,

(1963), 478–494.

[28] V. Lucarini and K.-E. Peiponen: Verification of generalized Kramers-Kronig relations and

sum rules on experimental data of third harmonic generation susceptibility on polymer, J.

Phys. Chem. 119 (2003), 620-627.

[29] F. Bassani and M. Altarelli: Interaction of radiation with condensed matter, in Handbook on

Synchrotron Radiation, E. E. Koch ed. (North Holland, Amsterdam, 1983)

[30] D. E. Aspnes: The accurate determination of optical properties by ellipsometry, in Handbook

of Optical Constants of Solids, E. D. Palik, ed. (Academic, New York, 1985), pp. 89-112

[31] K.-E. Peiponen and E. M. Vartiainen: Kramers-Kronig relations in optical data inversion,

Phys. Rev. B 44 (1991), 8301–8303.

[32] F. W. King: Efficient numerical approach to the evaluation of Kramers-Kronig transforms, J.

Opt. Soc. Am. B 19 (2002), 2427–2436.

[33] K. F. Palmer, M. Z. Williams, and B. A. Budde: Multiply subtractive Kramers-Kronig analysis

of optical data, Appl. Opt. 37 (1998), 2660–2673.

[34] V. Lucarini, J. J. Saarinen, and K.-E. Peiponen: Multiply subtractive generalized Kramers-

Kronig relations: Application on third-harmonic generation susceptibility on polysilane, J.

Chem. Phys. 119 (2003), 11095–11098.

[35] C. E. Leith: Climate response and fluctuation dissipation, J. Atmos. Sci. 32 (1975), 2022-2026.

[36] K. Lindenberg, B.J. West: Fluctuation and Dissipation in a Barotropic Flow Field, J. Atmos.

Sci. 41 (1984), 3021–3031.

[37] S. Corti, F. Molteni, T. N. Palmer: Signature of recent climate change in frequencies of natural

atmospheric circulation regimes, Nature 398, 799-802.

[38] V. Lucarini, A. Speranza, and R. Vitolo: Parametric smoothness and self-scaling of the statis-

tical properties of a minimal climate model: What beyond the mean field theories?, Physica

D 234 (2007), 105–123.

[39] V. Lucarini: Towards a definition of climate science, Int. J. Environ. Pollut. 18 (2002), 409–

414.

[40] A. Speranza, V. Lucarini: Environmental Science: physical principles and applications, in

Encyclopedia of Condensed Matter Physics, F. Bassani, J. Liedl, P. Wyder eds., Elsevier,

Amsterdam, (2005).

[41] Intergovernmental Panel on Climate Change, Climate Change 2007: The Physical Science

Basis. Contribution of Working Group I to the Fourth Assessment Report of the Intergovern-

mental Panel on Climate Change, Cambridge University Press, Cambridge, 2007.